# RLBench (/zh/docs/evaluation/benchmark-hub/rlbench)



## 简介 [#简介]

RLBench external benchmark task manifest for CoppeliaSim/PyRep robot manipulation rollouts.

## 指标 [#指标]

RLBench external benchmark task manifest for CoppeliaSim/PyRep robot manipulation rollouts.

### Task YAML [#task-yaml]

`worldfoundry/data/benchmarks/tasks/external/rlbench.yaml`

### Primary [#primary]

* `success_rate`

| Metric              | 关注点                                         |
| ------------------- | ------------------------------------------- |
| `success_rate`      | episode / task 是否成功完成的比例或成功率。               |
| `task_success`      | episode / task 是否成功完成的比例或成功率。               |
| `episode_success`   | episode / task 是否成功完成的比例或成功率。               |
| `reward`            | RLBench 桌面操作 episode 内累计的稀疏或 shaped reward。 |
| `normalized_return` | RLBench return 相对示范或成功基准归一化到 \[0, 1]。       |

## 环境准备 [#环境准备]

### 运行环境 [#运行环境]

* **Conda 环境：** `needs-rlbench-pyrep-coppeliasim-env`。
* **单独配置：** 需要 benchmark 专用 conda 环境，不要与统一 runtime 混用。
* RLBench 需要 CoppeliaSim/PyRep 组合环境；当前只记录 env vars 和 task manifest，不启动 simulator。

### 测评数据 [#测评数据]

* RLBench is simulator-backed; WorldFoundry normalizes policy rollout results while CoppeliaSim/PyRep execution remains explicit.
* 待测模型输出目录：设置 `WORLDFOUNDRY_RLBENCH_RESULTS_PATH` 指向生成视频或 rollout artifact 根目录。

### Checkpoint 与资产 [#checkpoint-与资产]

* 该 benchmark 以官方结果归一化或仿真 rollout 为主；policy / simulator checkpoint 见 catalog `requires` 与 task YAML。

### 关键环境变量 [#关键环境变量]

* `WORLDFOUNDRY_RLBENCH_ROOT`
* `COPPELIASIM_ROOT`
* `WORLDFOUNDRY_RLBENCH_RESULTS_PATH`

### 环境检查 [#环境检查]

```bash
PYTHONPATH=. ${WORLDFOUNDRY_CONDA_ROOT}/worldplay/bin/python -m worldfoundry.evaluation.tasks.execution.orchestration.runtime_preflight --profile rlbench --manifest worldfoundry/data/benchmarks/runtime_profiles/official/rlbench.yaml --output-dir tmp/worldfoundry_preflight/rlbench --json
```

## 运行评测 [#运行评测]

准备好该 benchmark 页面列出的资产和候选产物后，使用下面的公开运行入口。

### 设置候选产物目录 [#设置候选产物目录]

```bash
export WORLDFOUNDRY_GENERATED_ARTIFACT_DIR=/path/to/generated/artifacts
```

### 导入已有官方结果 [#导入已有官方结果]

```bash
worldfoundry-eval zoo benchmark-run \
  --benchmark-id rlbench \
  --mode normalizer \
  --official-results-path /path/to/official/results-or-report \
  --generated-artifact-dir "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
  --output-dir tmp/rlbench/normalizer \
  --json
```

### 从生成产物重新打分 [#从生成产物重新打分]

```bash
worldfoundry-eval zoo benchmark-run \
  --benchmark-id rlbench \
  --mode official-run \
  --generated-artifact-dir "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
  --output-dir tmp/rlbench/official-run \
  --json
```

### 直接调用仓内 runner [#直接调用仓内-runner]

```bash
worldfoundry-eval zoo benchmark-run --benchmark-id rlbench --mode official-validation --official-results-path <official_results.json> --output-dir <out> --json
```

## 依赖与产物 [#依赖与产物]

### Inputs [#inputs]

* Official RLBench repository checkout
* CoppeliaSim and PyRep runtime
* RLBench task assets
* Benchmark-specific policy checkpoint

### Environment [#environment]

* `WORLDFOUNDRY_RLBENCH_RESULTS_PATH`

### Outputs [#outputs]

* `scorecard.json`
* `raw_results.jsonl`

[← 返回 Benchmark Hub](/zh/docs/evaluation/benchmark-hub)